Tldr: it's built using Dhingra et al.'s Tweet2Vec system - https://github.com/bdhingra/tweet2vec to convert tweets into word vectors and then looks at the cosine similarity to find similar tweets of yore